I. Introduction
Sentiment analysis, the task referring to the automatic determination of user opinion from text, has received increased attention in the past decade [1]–[4]. Much of the success of sentiment analysis techniques can be attributed to the rise of social media platforms, where millions of users share their opinions on a wide variety of subjects. The majority of sentiment analysis methods are aimed at aggregating opinions towards entities like movies, people, products, or companies. We refer to this well-known research area as external sentiment analysis, in which sentiment and textual polarity is calculated with respect to a specific external entity. In contrast, we define internal sentiment analysis as the study of the polarity of user text with respect to themselves, primarily concerned with statements of emotion and mental health [5]. In this paper, we strictly focus on internal sentiment analysis, specifically with the valence prediction of private journals in a mental health therapy setting. Our work partly aligns with previous research regarding emotion detection in text [6]–[8], a subtask of the field of affective computing and analysis, but unlike previous work, we focus on the expansion of valence categories in a mental health setting.